Served on MixpeekFaceVerified 2026-11-05
InsightFace buffalo_l (SCRFD-10GF + ArcFace ResNet50 WebFace600K)
insightface/buffalo_l
InsightFace buffalo_l (SCRFD-10GF + ArcFace ResNet50 WebFace600K) is a face recognition model. Served on Mixpeek, scaling to zero when idle.
A Ray Serve app in Mixpeek's inference registry hosts it. It starts at zero replicas and wakes on first use.
New opens the Designer in Studio. For an existing namespace, open it in Studio and use its Designer tab.
No demo for InsightFace buffalo_l (SCRFD-10GF + ArcFace ResNet50 WebFace600K) yet
A video, image, slideshow or playground goes here once we record one on our own data.
What it is good and bad at
- Best at
- Face detection, 5-point alignment and 512-d face identity vectors for finding the same person across images and video frames.
- Worst at
- Commercial use (license); tiny or heavily occluded faces below min_face_size; the parameter names do not match the weights: scrfd_2.5g and scrfd_10g both load SCRFD-10GF, and arcface_r100 / magface_r100 load a ResNet50 (w600k_r50), not an R100.
- Task
- Face
- Input
- Image, Video
- Output
- Boxes, Keypoints, Vector
- Use cases
- Face search
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
- image
- det_size
- 640x640
- align
- 5-point affine to 112x112
- video
- frames sampled by the extractor, optional dedup of the same face
Output
- boxes
- face bounding boxes with detection scores
- keypoints
- 5 landmarks per face (2d106 and 3d68 landmark models also in the pack)
- labels
- gender and age from the pack's genderage head
- vector
- dims
- 512
- distance
- cosine
- type
- dense
How Mixpeek wires it
- Extractors
face_identity_extractor@v1detection_model=scrfd_2.5g (default) or scrfd_10g; embedding fixed arcface_r100
- Feature URIs
mixpeek://face_identity_extractor@v1/insightface__arcface
- Vector indexes
face_identity_extractor_v1_embedding- Retriever stages
- feature_search
- Combines with
- image or video -> SCRFD boxes -> aligned 112x112 crops -> ArcFace 512-d vector per face -> feature_search with a face photo
How it runs
- Status
- Served on Mixpeek insightface__arcface Serve app; face_identity_extractor batch calls it (FaceProcessingBatchService, strict_serve, no local fallback).
- Serve apps
insightface__arcface- Runtime
- insightface.app.FaceAnalysis(name="buffalo_l") on onnxruntime, prepare(ctx_id from providers, det_size=(640, 640))
- Weights
- ONNX, no remote code
- Parameters
- Not recorded
- Size on disk
- 326 MB
- Preprocessing
- SCRFD detection at 640x640, 5-point affine alignment to 112x112.
- Postprocessing
- ArcFace recognition embedding per aligned face (512 dims), L2-normalized; optional quality scoring and video dedup.
- Serving
- both, batch 8
- Hardware
- Runs on CPU (slower without a GPU)
Scale to zero
- Minimum replicas
- 0
- Maximum replicas
- 4
- Idle timeout
- 600 s
- Memory
- 2 GB
- CPUs
- 0.5
- GPUs
- Set at deploy time: hardware_config.gpu_fraction if hardware_config.gpu_concurrency > 0 else 0
- Cold start, node warm
- Not recorded
- Cold start, new node
- Not recorded
Measured cost
Unmeasured
No committed measurement. The definition's detection latencies (2-3, 5-7, 10-15 ms) are descriptions, not committed measurements.
Our evaluations
Unmeasured
We have not evaluated this model on our own data yet. Publisher results are in the sources below.
Questions
- What is InsightFace buffalo_l (SCRFD-10GF + ArcFace ResNet50 WebFace600K) best at?
- Face detection, 5-point alignment and 512-d face identity vectors for finding the same person across images and video frames.
- What are the limits of InsightFace buffalo_l (SCRFD-10GF + ArcFace ResNet50 WebFace600K)?
- Commercial use (license); tiny or heavily occluded faces below min_face_size; the parameter names do not match the weights: scrfd_2.5g and scrfd_10g both load SCRFD-10GF, and arcface_r100 / magface_r100 load a ResNet50 (w600k_r50), not an R100.
- Can InsightFace buffalo_l (SCRFD-10GF + ArcFace ResNet50 WebFace600K) be used commercially?
- No. The insightface README and model_zoo state every pretrained model, downloaded automatically or manually, is for non-commercial research only; the MIT license covers the library code. face_identity_extractor bills customers for this pack, so a commercial license from InsightFace is needed.
- How does InsightFace buffalo_l (SCRFD-10GF + ArcFace ResNet50 WebFace600K) run on Mixpeek?
- insightface__arcface Serve app; face_identity_extractor batch calls it (FaceProcessingBatchService, strict_serve, no local fallback).